How can inference risk be reduced in AI?

Restrict inputs and add monitoring tools, Let users review and retrain the model

How are hardware attacks used in AI?

Disrupting systems with physical faults

What is a key goal of adversarial inputs?

To trick the model into misclassifying data

What is a prompt injection in LLMs?

Prompts crafted to override model intent

Which is a key part of adversarial strategy?

Sending inputs to confuse the model

What is data extraction in AI threats?

Copying a model’s behavior through many queries

What happens during a model inversion attack?

Sensitive training data is recovered by attackers

Which scenario is a poisoning attack example?

Embedding triggers in the training data

What is poisoning in machine learning?

Corrupting training data to shift outcomes

What defines a white-box attack in AI?

The attacker knows the model’s internal structure

Why is prompt injection a threat to LLMs?

To change model behavior using bad prompts

Why are AI models vulnerable during training?

Attackers can inject malicious data

What does "training phase" primarily refer to?

Building and validating the AI model

What makes APIs risky in AI systems?

They expose the model to external inputs

What does the inference stage in AI typically involve?

Generating outputs from a trained model

What is a common mitigation for AI risk?

Limit access and monitor input behavior

What’s the focus of CRISP-ML(Q)?

Secure every phase of AI development

What makes AI threats different from normal cyberattacks?

They target learning stages and input data

Why are training and inference both secured?

Both stages can be exploited differently

What is the MITRE ATT&CK framework for?

Describing tactics used in cyberattacks

What do fault injection attacks involve?

Using physical or data glitches to disrupt

What does CRISP-ML(Q) help organizations achieve?

Apply secure steps through model lifecycles

studbud

INFT 1204 FINAL

25 termsStudBud Administration29 plays

Imported from external platform

Flashcards25 cardsFPractice quizPHost liveL

Questions 25

Answers

1
Which option is a typical AI access point?
Wireless keyboards
User interfaces
SSD bootloaders
USB debugging ports
2
How can inference risk be reduced in AI?
Restrict inputs and add monitoring tools
Bypass model outputs to stop bad actors
Let users review and retrain the model
Rely on memory-free physical deployment
3
How are hardware attacks used in AI?
Encrypting data with external chip readers
Disrupting systems with physical faults
Sending queries to clean input layers
Opening virtual sessions via side ports
4
What is a key goal of adversarial inputs?
To overload the GPU
To reduce accuracy intentionally
To trick the model into misclassifying data
To trigger software updates
5
What is a prompt injection in LLMs?
Training logs that leak into prompts
Prompts crafted to override model intent
Injecting input from cache in offline mode
Prompts skipped due to low confidence
6
Which is a key part of adversarial strategy?
Making logs harder for devs to analyze
Sending inputs to confuse the model
Replacing APIs with read-only endpoints
Upgrading the model without approval
7
What is data extraction in AI threats?
Hiding models inside software containers
Using random prompts to confuse the model
Writing backup copies on external drives
Copying a model’s behavior through many queries
8
What happens during a model inversion attack?
Encryption keys are swapped during backups
Log files are removed before testing begins
Sensitive training data is recovered by attackers
The training data and model are compressed and hashed
9
Which scenario is a poisoning attack example?
Adding training dataset confusion using runtime monitoring
Embedding triggers in the training data
Uploading text prompts via VPN tunnels
Deleting old data from training logs and model
10
What is poisoning in machine learning?
Corrupting training data to shift outcomes
Disabling log systems to avoid alerts
Restoring models with backup configurations
Replacing output values with static tags
11
The aesthetics of the user interface is NOT key ML attack factor
False
True
12
What defines a white-box attack in AI?
The attacker interacts only through inputs and outputs
The attacker exploits firewall misconfigurations
The attacker modifies hardware interfaces
The attacker knows the model’s internal structure
13
Why is prompt injection a threat to LLMs?
To reroute requests to offline processors
To install scripts into system log files
To delete queries stored in cache memory
To change model behavior using bad prompts
14
Why are AI models vulnerable during training?
Attackers can inject malicious data
The model uses outdated logic
They have limited access to inference data
They rely on black-box system knowledge
15
What does "training phase" primarily refer to?
Buiding reports from AI model inference results
Downloading old test case libraries
Encrypting the system for validation
Building and validating the AI model
16
What makes APIs risky in AI systems?
They expose the model to external inputs
They block AI application access to memory logs
They can’t run while the model trains
They copy model responses to system backups
17
What does the inference stage in AI typically involve?
Collecting and labeling raw data
Cleaning and storing training samples
Developing security protocols for APIs
Generating outputs from a trained model
18
What is a common mitigation for AI risk?
Limit access and monitor input behavior
Store AI backups inside open file systems
Let users access logs and retrain models
Let models train with live production data
19
What’s the focus of CRISP-ML(Q)?
Improve visuals for web model demos
Secure every phase of AI development
Automate data prep using cloud scripts
Write AI output into secure databases
20
What makes AI threats different from normal cyberattacks?
They target learning stages and input data
They work only on encrypted traffic streams
They use Wi-Fi exploits to reach AI devices
They attack AI servers and input fake patches
21
Why are training and inference both secured?
Training runs longer and is more costly
Both stages can be exploited differently
Logs are easier to read while training
Inference skips encryption entirely
22
What is the MITRE ATT&CK framework for?
Building trustworthy large-scale AI private data sets
C. Describing automated malware payloads
Hiding AI models behind a firewall
Describing tactics used in cyberattacks
23
What do fault injection attacks involve?
Deleting AI updates from firmware
Logging prompt responses for tuning
Using physical or data glitches to disrupt
Querying models with outdated commands
24
What is meant by system knowledge?
Caching answers in encrypted files
Replacing labels in test datasets
Knowing how the AI model works
Knowing how the AI prompts works based on usage
25
What does CRISP-ML(Q) help organizations achieve?
Remove training steps to boost speed
Optimize ML for training simulations
Align LLMs lifecycle with cloud computing costs
Apply secure steps through model lifecycles

INFT 1204 FINAL Flashcards | StudBud